Reference Issues/PRs

What does this implement/fix? Explain your changes.

fix：#20589
PCA returns highly inaccurate results when number of features is large #20589

Any other comments?

The variable n_oversamples in func sklearn.utils.extmath.randomized_svd is 10 by default. The outer function sklearn.decomposition.PCA.fit_transform cannot modify the variable. So when the input data feature is greater than 10, the result of svd will be quite different and there will be a big error.
In order to solve this problem, I provided a sampling ratio(n_oversamples_rate) as an input parameter, and the user can customize the required error range according to the needs of the project. When the n_oversamples_rate=1, the svd result will be the same as scipy.linalg.svd and the scipy.sparse.linalg.svds.

@jeremiedbb you probably want to give your advise regarding which API to use for this parameter.

I'd keep the API of randomized_svd, i.e an integer. I agree that it might be interesting to have it being a float representing a ratio instead of an absolute value. But then we should discuss about changing it in randomized_svd directly.

@jeremiedbb you probably want to give your advise regarding which API to use for this parameter.

I'd keep the API of randomized_svd, i.e an integer. I agree that it might be interesting to have it being a float representing a ratio instead of an absolute value. But then we should discuss about changing it in randomized_svd directly.

First，thank you for your reply.

1. Integer is simpler, because we can know the dimension of the feature immediately, but float may be more convenient to use in the engineering, because we can determine the amount of features that need to be retained according to the needs. Sometimes we don’t really care about the specific dimension, we are more concerned about the feature retention;
   Finally, the two have their own advantages and disadvantages, or we can keep both at the same time, and users can customize their choices according to their needs.
2. However, no matter what kind of plan, we have to expose this parameter to users anyway, let them determine this parameter.
   At last, because I am also full of interest in machine learning, I hope my research can have a better effect on the improvement of the method.
   Thanks.

@x-shadow-man: the CI fails because of some changes made to the code have not been properly formatted.

Setting up the pre-commit setup (see the optional step n.9 of this section) will make you able to work more easily and will prevent problems in your code and in the jobs run on the CI. 🙂

You should check that my suggestion are working with black

Thanks @x-shadow-man Merging

Thanks for your help, I am also happy to solve this problem!